The Google Brain and Google AI team members have open-sourced the AI tool EfficientDet, a collection of advanced object detection models. The tool is an upgraded version of EfficientNet, with additional codes, that was launched last year for Coral Boards. 

Creators of the system, Google engineers Mingxing Tan, Google Ruoming Pang, and Quoc Le, also claim that EfficientDet achieves state-of-the-art object detection at a faster rate than other popular object detection models such as YOLO and AmoebaNet, when used with CPUs or GPUs. EfficientDet also displayed exceptional performance when it came to semantic segmentation experiments that were conducted with the PASCAL visual object challenge dataset. In EfficientDet, a bidirectional feature pyramid network (BiFPN) acts as a feature network, and ImageNet pre-trained EfficientNet acts as the backbone network.

“Aiming at optimizing both accuracy and efficiency, we would like to develop a family of models that can meet a wide spectrum of resource constraints,” according to the paper titled EfficientDet: Scalable and efficient object detection, which examines neural network architecture design for object detection. The existing object detection tools are either inaccurate or are extremely resource-intensive. EfficientDet, however, is less expensive and achieves quality with a method that “uniformly scales the resolution, depth, and width for all backbone, feature network, and box/class prediction networks at the same time.”

“The large model sizes and expensive computation costs deter their deployment in many real-world applications such as robotics and self-driving cars where model size and latency are highly constrained,” the paper reads. “Given these real-world resource constraints, model efficiency becomes increasingly important for object detection.” EfficientDet optimizes for cross-scale connections in part by removing nodes that only have one input edge to create a simpler bidirectional network. It also relies on the one-stage detector paradigm, an object detector known for efficiency and simplicity.

“We propose to add an additional weight for each input during feature fusion and let the network to learn the importance of each input feature,” the paper reads.